講演情報
[16p-K306-6]Simulation of In-materio Physical Reservoir using Unidirectional Cluster-Based Echo State Network
〇(M1)Tu Truong Huynh1, Moulika Desu1, Yuichiro Tanaka1,2, Hirofumi Tanaka1,2 (1.LSSE, Kyutech, 2.Neumorph Center, Kyutech)
キーワード：
in-materio physical reservoir simulation、Echo State Network、reservoir computing
In fabricated in-materio physical reservoir computing devices, mapping their internal network structure poses a significant challenge due to variability in bonding, size, and materials. Simulating these networks is essential to understand and investigate how these devices operate. Echo state network (ESN) provides a promising framework as it can mimic the nonlinear and transient behaviors observed in physical devices. However, conventional ESN simulations take computational time, especially when processing large networks, and are often prone to encountering inconsistent outputs due to random characteristics. They also produce unsuitable results that are not similar to real-life devices. Therefore, we propose introducing a constraint into ESN simulations that allows just neighbor nodes to connect. The constraint comes from an assumption that only neighbor nodes in in-materio reservoir affect each other, and it can reduce the cost of computation and hyperparameter tuning. Our goal is to develop a CPU-optimized ESN simulation framework that (1) closely matches real-life output, (2) delivers highly consistent, reproducible results, (3) requires minimal to no hyperparameter adjustment, and (4) operates at least 30% faster than conventional ESNs.